Abstract
In this paper, we proposed a method to generate synthetic brain images using generative adversarial networks (GANs). In medical image analysis, it remains a difficult and important task to produce realistic medical images that are entirely diffe rent from the original ones and also the exchange of clinical image data is a crucial issue for the implementation of diagnostic support systems. Nonetheless, it is difficult for researchers to obtain medical image data because the images contain individual information. Recently proposed GAN models could learn how to distribute training images without seeing actual image data, and generated images can completely anonymized personal information. The produced images can be used as training images for the classification of medical image, promoting medical image analysis viable. Instead of collecting a large amount of MRI data, an approach to image generation has been implemented in our paper. We exploit a progressive growing GAN (PGGAN), a neonatal brain image generation method that can be used for brain MRI classification and ADHD prediction tasks. The PGGAN slowly discovers the features of ADHD in MRI images by adding new layers during the training phase. Our image generation approach shows that it can produce brain MR images avoid ing artificial artifacts and have clinical characteristics of the target symptom.
